CN113344222A - Safe and credible federal learning mechanism based on block chain - Google Patents

Abstract

The invention relates to a safe and credible federal learning mechanism based on a block chain.A participant carries out agency re-encryption on a shared secret key, the participant of a federal task can obtain an agency re-encryption ciphertext only by verifying the identity and writing the identity information into an access strategy, and the uploading and downloading parameters of the participant in the federal learning are realized in a transaction mode by utilizing the decentralized characteristic of the block chain, so that each participant has the obligation of maintaining a global federal learning model, the trust relationship among the participants is increased, the safety of user data is protected by adopting the technical scheme, and the fairness of the learning mechanism is ensured.

Description

Safe and credible federal learning mechanism based on block chain

Technical Field

The invention relates to the technical field of federal learning, in particular to a safe and credible federal learning mechanism based on a block chain.

Background

In a big data environment, the security and privacy problems of user data are receiving wide attention. The occurrence of federal learning reduces the risk of privacy disclosure of local sensitive data on one hand, and solves the problem that access limited data are difficult to be processed in a centralized manner by a server on the other hand. The concept of federal learning was originally proposed by google in 2016, and was originally used to solve the problem of local model update of android phone end users, but nowadays, the application of federal learning is also very common in our daily lives.

However, federal learning still presents a number of problems. First, the method of directly sharing parameters in federal learning does not completely protect data privacy. The unprocessed parameters are directly transmitted in a channel, once the parameters are intercepted, an attacker can restore partial data through the parameters, and the privacy of a user still has a leakage risk. Finally, since model sharing can reveal training data information, if the server is curious, even if a small portion of the gradient is uploaded, user privacy can be indirectly revealed.

Secondly, each participant in federal learning can access the global model, so that the possibility that a malicious participant adds or tampers with data weights to the global model in the training stage is very high, and once the weight parameters are tamped, the training data sets or local models of other participants are directly influenced, and further the performance and accuracy of the global model are indirectly influenced. Moreover, some participants may wish to benefit from the benefits of the global model while circumventing the risk of data leakage. The method does not use a local data training model, selects methods such as inserting virtual updating or not uploading the local model and the like to maintain connection with the environment, does not contribute to the training process, and finally causes the lack of fairness of the whole federal learning mechanism and low enthusiasm of participants.

In summary, in the existing federal learning mechanism, the user data security has a leakage risk, and the fairness of the learning mechanism is difficult to guarantee.

Disclosure of Invention

Aiming at the defects of the prior art, the invention provides a safe and credible federal learning mechanism based on a block chain, which prevents authorization infringement and has a fair learning mechanism.

The invention relates to a safe and credible federal learning mechanism based on a block chain, which comprises the following steps:

s1: the participator issues the federal learning task in the block chain and applies for the federal learning taskSymmetric key EK^kThen, for this symmetric key EK^kCarrying out proxy re-encryption to obtain proxy re-encrypted ciphertext, and decrypting the proxy re-encrypted ciphertext by the private key of the participant of the federal learning task to obtain the symmetric key EK^kBy means of a symmetric key EK^kDecrypting to obtain a federal learning task;

s2: after completing the federal learning task, the participants of the federal learning task send requests for uploading local model parameters W to the block chain links, the block chain links verify the identity information of the local model parameters, and the identity information is verified through a symmetric key EK^kEncrypting and uploading local model parameters to a server and obtaining point reward V_upForming a chain of upload transaction bill write blocks, and the server updates the global model parameter W according to the upload local model parameter_new；

S3: participants of the federal learning task send download global model parameters W_newTo block link points, which verify their identity information, and the participant spending point V of the federal learning task_dowForming a chain of write blocks for downloading transaction bills, downloading the encrypted ciphertext from the server, and using the symmetric key EK^kDecrypting the encrypted ciphertext to obtain global model parameter W_new。

Further, the specific step of step S1 includes:

s1-1: participant issues kth task request in blockchain

And applies for the symmetric key EK of the task to the key center^kWhile using the public key of the participant to the symmetric key EK of the task^kCarrying out re-encryption to obtain a re-encrypted ciphertext and sending the re-encrypted ciphertext to the cloud server;

s1-2: the other participants receive the kth task request

Selecting whether to join Federal learning task T^kIf the identity information and the public key thereof are selected to be added, the identity information and the public key thereof are written into an access POLICY^k；

S1-3: participant verification visitChallenge strategy POLICY^kGenerating a proxy key by using a public key of a participant added into the federal learning task through a proxy key algorithm RKGen (), and sending the proxy key to a server;

s1-4: after obtaining the proxy key, the server reads the re-encrypted ciphertext, converts the re-encrypted ciphertext into a proxy re-encrypted ciphertext by using the proxy key through a re-encryption algorithm ReEncrypt (), stores the proxy re-encrypted ciphertext, and returns a position index including a proxy re-encrypted ciphertext storage position to be stored in a block chain;

s1-5: the participants of the federal learning task obtain the position index from the block chain, download the proxy re-encrypted ciphertext from the server, decrypt the proxy re-encrypted ciphertext by using the private key of the proxy re-encrypted ciphertext to obtain the symmetric key EK^k；

S1-6: participant to federal learning tasks uses EK^kAnd decrypting to obtain the federal learning task.

Further, the specific step of step S2 includes:

s2-1: participants of federated learning tasks are directed to task requests

In the q-th turn, after completing the federal learning task, sending a request for uploading a local model parameter W to the block chain link;

s2-2: the block link points broadcast the validity of the signature of the participant of the verification federal learning task to the inner nodes; if the verification is successful, executing S2-3; if the verification fails, the feedback request fails;

s2-3: the block chain node temporarily stores the request into a local log of the block chain node and waits for other nodes to respond; after more than half of nodes in the block chain agree, sending a message of successful feedback request and a signature set Sign agreed by the nodes to participants of the federal learning task, and executing S2-4; otherwise, the feedback request fails;

s2-4: the participants of the federal learning task receive the feedback message, upload the local model parameters and use the symmetric key EK^kEncrypting G the learning efficiency-alpha and the local gradient of the negative value to be uploaded to obtain a ciphertext E^k(－α·G)；And the signature Sign { E obtained thereby^kUploading to a cloud server and obtaining point reward, which is marked as V_upSimultaneously broadcasting the transaction bill to the block chain; wherein the uploaded transaction bill may be recorded as:

wherein, T_{k_q}Representing the q-th turn of the k-th federal learning task for the task identification;

representing the serial number of the uploaded transaction bill; t is t_0→iIs a time stamp of the transaction bill; id₀Is the identity id of the participant;

is the content signed by a private key when uploading by a participant; l is_iA parameter location index representing the ith upload of the participant; v_upRepresenting the amount of the bonus points earned by the participant; b is₀Representing the account balance of the participant; a isⁱRepresenting the number of times that the current round participant is downloaded by other participants when uploading the local parameters for the ith time; tranN_iA trade order number representing the trade download of the other participant;

s2-5: the cloud server receives the ciphertext E^k(. alpha. -G) and signature Sign { E)^k(- α · G) }, verifying its signature; if the authentication is successful, executing S2-6, and if the authentication is failed, giving up the update;

s2-6: the server receives the ciphertext E based on the principle of addition homomorphic encryption^kNew model parameter W of (-alpha G) addition calculation_newAnd store the ciphertext of:

E^k(W_new)＝E^k(W_old+(－α·G))＝E^k(W_old)+E^k(－α·G)

wherein W_newRepresenting new model parameters, W_oldRepresenting the model parameters to be updated, E^k() represents a homomorphic encrypted ciphertext; then theReturning a message containing ciphertext E^k(W_new) The location of the storage location is indexed to the blockchain.

Further, the specific step of step S3 includes:

s3-1: participants of the federal learning task send download global model parameters W_newTo block link points;

s3-2: the block chain node needs to broadcast and verify the identity of the participants of the federal learning task and check the participants and the task T^kAccess POLICY of^kWhether the identity information in (1) is consistent; if yes, executing S3-3; if not, the feedback downloading fails;

s3-3: participant of federated learning task obtains ciphertext E in blockchain^k(W_new) Storing the position index of the position, and downloading the ciphertext E from the server^k(W_new) Simultaneously broadcasting and downloading the transaction bill to the block chain; wherein, the download transaction bill can be recorded as:

wherein, T_{k_q}Representing the q-th turn of the k-th federal learning task for the task identification;

indicating the download transaction bill number, t_0→iIs a time stamp of the transaction bill; id_iIs the identity id of the participant;

is the content signed by the private key when the participant downloads; l is_iRepresenting a parameter location index; v_dowA credit amount representing a download cost; b is_iRepresenting the credit balance of the participant; b is₀Representing the current point balance of the uploading participant in the round;

s3-4: symmetric key EK is used by participants in the federal learning task^kDecrypting ciphertext E^k(W_new) Obtaining global model parameters W_new。

Compared with the prior art, the invention has the following beneficial effects:

1. the participants perform proxy re-encryption on the shared secret key, and the participants of the federal task can obtain the proxy re-encryption ciphertext only by verifying the identity and writing the identity information into an access strategy. Therefore, dynamic management of shared keys is realized, and the keys of all tasks in federated learning are different, so that the federated learning tasks are independent from one another, and participants of the tasks are independent from one another.

2. By utilizing the decentralized characteristic of the block chain, uploading and downloading parameters of participants in the federal learning are realized in a trading mode, so that each participant has the obligation of maintaining a global federal learning model, and the trust relationship among the participants is increased. And the block chain based consensus mechanism manages account transactions from two aspects. On one hand, the participants who upload the parameters are rewarded with basic money, and on the other hand, the downloading participants need to pay a certain amount of money to the current round of the uploaders to download. The mechanism is operated based on the block chain, and is maintained and executed by participants, so that the fairness of the operation of the mechanism is ensured.

Drawings

The accompanying drawings, which are described herein to provide a further understanding of the application, are included in the following description:

FIG. 1 is a flow chart of the present invention;

FIG. 2 is a flow chart of step two of the present invention;

FIG. 3 is a flow chart of step three of the present invention.

Detailed Description

Referring to fig. 1, in an embodiment, the invention relates to a safe and reliable federal learning mechanism based on a block chain, which includes the following steps:

s1: the participator issues the federal learning task in the block chain and applies for the symmetric key EK of the federal learning task^kThen, for this symmetric key EK^kCarrying out proxy re-encryption to obtain proxy re-encrypted ciphertext, and decrypting the proxy re-encrypted ciphertext by the private key of the participant of the federal learning task to obtain the symmetric key EK^kBy means of a symmetric key EK^kDecrypting to obtain a federal learning task;

s2: after completing the federal learning task, the participants of the federal learning task send requests for uploading local model parameters W to the block chain links, the block chain links verify the identity information of the local model parameters, and the identity information is verified through a symmetric key EK^kEncrypting and uploading local model parameters to a server and obtaining point reward V_upForming a chain of upload transaction bill write blocks, and the server updates the global model parameter W according to the upload local model parameter_new；

S3: participants of the federal learning task send download global model parameters W_newTo block link points, which verify their identity information, and the participant spending point V of the federal learning task_dowForming a chain of write blocks for downloading transaction bills, downloading the encrypted ciphertext from the server, and using the symmetric key EK^kDecrypting the encrypted ciphertext to obtain global model parameter W_new。

Of course, the participants of the federated learning task may be based on the global model parameters W_newAnd modifying the local model parameters, uploading again and executing S2 until the requirements of the participants are met, and ending the federal learning task.

Further specific embodiments of step S1 include:

s1-1: participant issues kth task request in blockchain

And applies for the symmetric key EK of the task to the key center^kWhile using the public key of the participant to the symmetric key EK of the task^kCarrying out re-encryption to obtain a re-encrypted ciphertext and sending the re-encrypted ciphertext to the cloud server;

s1-2: the other participants receive the kth task request

Selecting whether to join Federal learning task T^kIf the identity information and the public key thereof are selected to be added, the identity information and the public key thereof are written into an access POLICY^k；

S1-3: participant authentication access policiesSlightly POLICY^kGenerating a proxy key by using a public key of a participant added into the federal learning task through a proxy key algorithm RKGen (), and sending the proxy key to a server;

s1-4: after obtaining the proxy key, the server reads the re-encrypted ciphertext, converts the re-encrypted ciphertext into a proxy re-encrypted ciphertext by using the proxy key through a re-encryption algorithm ReEncrypt (), stores the proxy re-encrypted ciphertext, and returns a position index including a proxy re-encrypted ciphertext storage position to be stored in a block chain;

s1-5: the participants of the federal learning task obtain the position index from the block chain, download the proxy re-encrypted ciphertext from the server, decrypt the proxy re-encrypted ciphertext by using the private key of the proxy re-encrypted ciphertext to obtain the symmetric key EK^k；

S1-6: participant to federal learning tasks uses EK^kAnd decrypting to obtain the federal learning task.

Referring to fig. 2, the further specific step of step S2 includes:

s2-1: participants of federated learning tasks are directed to task requests

In the q-th turn, after completing the federal learning task, sending a request for uploading a local model parameter W to the block chain link;

s2-2: the block link points broadcast the validity of the signature of the participant of the verification federal learning task to the inner nodes; if the verification is successful, executing S2-3; if the verification fails, the feedback request fails, and the request for uploading the local model parameter W is finished;

s2-3: the block chain node temporarily stores the request into a local log of the block chain node and waits for other nodes to respond; after more than half of nodes in the block chain agree, sending a message of successful feedback request and a signature set Sign agreed by the nodes to participants of the federal learning task, and executing S2-4; otherwise, the feedback request fails; otherwise, the feedback request fails, and the request for uploading the local model parameter W is finished;

s2-4: participants of the federal learning task receive the feedback message, upload local model parameters, and use symmetryKey EK^kEncrypting G the learning efficiency-alpha and the local gradient of the negative value to be uploaded to obtain a ciphertext E^k(-. G); and the signature Sign { E obtained thereby^kUploading to a cloud server and obtaining point reward, which is marked as V_upSimultaneously broadcasting the transaction bill to the block chain; wherein the uploaded transaction bill may be recorded as:

wherein, T_{k_q}Representing the q-th turn of the k-th federal learning task for the task identification;

representing the serial number of the uploaded transaction bill; t is t_0→iIs a time stamp of the transaction bill; id₀Is the identity id of the participant;

is the content signed by a private key when uploading by a participant; l is_iA parameter location index representing the ith upload of the participant; v_upRepresenting the amount of the bonus points earned by the participant; b is₀Representing the account balance of the participant; a isⁱRepresenting the number of times that the current round participant is downloaded by other participants when uploading the local parameters for the ith time; tranN_iA trade order number representing the trade download of the other participant;

s2-5: the cloud server receives the ciphertext E^k(. alpha. -G) and signature Sign { E)^k(- α · G) }, verifying its signature; executing S2-6 if the identity authentication is successful, giving up updating if the identity authentication is failed, and ending the request of uploading the local model parameter W;

s2-6: the server receives the ciphertext E based on the principle of addition homomorphic encryption^kNew model parameter W of (-alpha G) addition calculation_newAnd store the ciphertext of:

E^k(W_new)＝E^k(W_old+(－α·G))＝E^k(W_old)+E^k(－α·G)

wherein W_newRepresenting new model parameters, W_oldRepresenting the model parameters to be updated, E^k() represents a homomorphic encrypted ciphertext; then returns a message containing the ciphertext E^k(W_new) And the position index of the storage position is indexed to the block chain, and the request for uploading the parameter W is completed.

Referring to fig. 3, a further specific embodiment of the step S3 includes:

s3-1: participants of the federal learning task send download global model parameters W_newTo block link points;

s3-2: the block chain node needs to broadcast and verify the identity of the participants of the federal learning task and check the participants and the task T^kAccess POLICY of^kWhether the identity information in (1) is consistent; if yes, executing S3-3; if not, the feedback downloading fails, and the downloading of the local model parameter W is finished_newA request for (2);

s3-3: participant of federated learning task obtains ciphertext E in blockchain^k(W_new) Storing the position index of the position, and downloading the ciphertext E from the server^k(W_new) Simultaneously broadcasting and downloading the transaction bill to the block chain; wherein, the download transaction bill can be recorded as:

wherein, T_{k_q}Representing the q-th turn of the k-th federal learning task for the task identification;

indicating the download transaction bill number, t_0→iIs a time stamp of the transaction bill; id_iIs the identity id of the participant;

is the content signed by the private key when the participant downloads; l is_iRepresenting a parameter location index; v_dowCredit amount indicating download cost；B_iRepresenting the credit balance of the participant; b is₀Representing the current point balance of the uploading participant in the round;

s3-4: symmetric key EK is used by participants in the federal learning task^kDecrypting ciphertext E^k(W_new) Obtaining global model parameters W_newCompleting the local model parameter W downloaded this time_newThe request of (1).

Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that changes, modifications, substitutions and alterations can be made in these embodiments without departing from the principles and spirit of the invention, the scope of which is defined in the appended claims and their equivalents.

Claims (4)

1. A safe and credible federal learning mechanism based on a block chain is characterized in that:

the method comprises the following steps:

s1: the participator issues the federal learning task in the block chain and applies for the symmetric key EK of the federal learning task^kThen, for this symmetric key EK^kCarrying out proxy re-encryption to obtain proxy re-encrypted ciphertext, and decrypting the proxy re-encrypted ciphertext by the private key of the participant of the federal learning task to obtain the symmetric key EK^kBy means of a symmetric key EK^kDecrypting to obtain a federal learning task;

s2: after completing the federal learning task, the participants of the federal learning task send requests for uploading local model parameters W to the block chain links, the block chain links verify the identity information of the local model parameters, and the identity information is verified through a symmetric key EK^kEncrypting and uploading local model parameters to a server and obtaining point reward V_upForming a chain of upload transaction bill write blocks, and the server updates the global model parameter W according to the upload local model parameter_new；

S3: participants of the federal learning task send download global model parameters W_newTo block link points, which verify their identity information, and the participant spending point V of the federal learning task_dowForming download transaction bill writesBlock chaining and downloading encrypted ciphertext from a server using a symmetric key EK^kDecrypting the encrypted ciphertext to obtain global model parameter W_new。

2. The mechanism of claim 1, wherein the mechanism is characterized in that:

the specific steps of step S1 include:

s1-1: participant issues kth task request in blockchain

And applies for the symmetric key EK of the task to the key center^kWhile using the public key of the participant to the symmetric key EK of the task^kCarrying out re-encryption to obtain a re-encrypted ciphertext and sending the re-encrypted ciphertext to the cloud server;

s1-2: the other participants receive the kth task request

Selecting whether to join Federal learning task T^kIf the identity information and the public key thereof are selected to be added, the identity information and the public key thereof are written into an access POLICY^k；

S1-3: participant authentication access POLICY^kGenerating a proxy key by using a public key of a participant added into the federal learning task through a proxy key algorithm RKGen (), and sending the proxy key to a server;

s1-4: after obtaining the proxy key, the server reads the re-encrypted ciphertext, converts the re-encrypted ciphertext into a proxy re-encrypted ciphertext by using the proxy key through a re-encryption algorithm ReEncrypt (), stores the proxy re-encrypted ciphertext, and returns a position index including a proxy re-encrypted ciphertext storage position to be stored in a block chain;

s1-5: the participants of the federal learning task obtain the position index from the block chain, download the proxy re-encrypted ciphertext from the server, decrypt the proxy re-encrypted ciphertext by using the private key of the proxy re-encrypted ciphertext to obtain the symmetric key EK^k；

S1-6: participant to federal learning tasks uses EK^kAnd decrypting to obtain the federal learning task.

3. The mechanism of claim 1, wherein the mechanism is characterized in that:

the specific steps of step S2 include:

s2-1: participants of federated learning tasks are directed to task requests

In the q-th turn, after completing the federal learning task, sending a request for uploading a local model parameter W to the block chain link;

s2-2: the block link points broadcast the validity of the signature of the participant of the verification federal learning task to the inner nodes; if the verification is successful, executing S2-3; if the verification fails, the feedback request fails;

s2-3: the block chain node temporarily stores the request into a local log of the block chain node and waits for other nodes to respond; after more than half of nodes in the block chain agree, sending a message of successful feedback request and a signature set Sign agreed by the nodes to participants of the federal learning task, and executing S2-4; otherwise, the feedback request fails;

s2-4: the participants of the federal learning task receive the feedback message, upload the local model parameters and use the symmetric key EK^kEncrypting G the learning efficiency-alpha and the local gradient of the negative value to be uploaded to obtain a ciphertext E^k(-. G); and the signature Sign { E obtained thereby^kUploading to a cloud server and obtaining point reward, which is marked as V_upSimultaneously broadcasting the transaction bill to the block chain; wherein the uploaded transaction bill may be recorded as:

wherein, T_{k_q}Representing the q-th turn of the k-th federal learning task for the task identification;

representing the serial number of the uploaded transaction bill; t is t_0→iIs a time stamp of the transaction bill; id₀Is the identity id of the participant;

is the content signed by a private key when uploading by a participant; l is_iA parameter location index representing the ith upload of the participant; v_upRepresenting the amount of the bonus points earned by the participant; b is₀Representing the account balance of the participant; a isⁱRepresenting the number of times that the current round participant is downloaded by other participants when uploading the local parameters for the ith time; tranN_iA trade order number representing the trade download of the other participant;

s2-5: the cloud server receives the ciphertext E^k(. alpha. -G) and signature Sign { E)^k(- α · G) }, verifying its signature; if the authentication is successful, executing S2-6, and if the authentication is failed, giving up the update;

s2-6: the server receives the ciphertext E based on the principle of addition homomorphic encryption^kNew model parameter W of (-alpha G) addition calculation_newAnd store the ciphertext of:

E^k(W_new)＝E^k(W_old+(－α·G))＝E^k(W_old)+E^k(－α·G)

wherein W_newRepresenting new model parameters, W_oldRepresenting the model parameters to be updated, E^k() represents a homomorphic encrypted ciphertext; then returns a message containing the ciphertext E^k(W_new) The location of the storage location is indexed to the blockchain.

4. The mechanism of claim 1, wherein the mechanism is characterized in that:

the specific steps of step S3 include:

s3-1: participants of the federal learning task send download global model parameters W_newTo block link points;

s3-2: block link point needs to be wideBroadcasting verification of the identity of the participants of the federal learning task, checking it with the task T^kAccess POLICY of^kWhether the identity information in (1) is consistent; if yes, executing S3-3; if not, the feedback downloading fails;

s3-3: participant of federated learning task obtains ciphertext E in blockchain^k(W_new) Storing the position index of the position, and downloading the ciphertext E from the server^k(W_new) Simultaneously broadcasting and downloading the transaction bill to the block chain; wherein, the download transaction bill can be recorded as:

wherein, T_{k_q}Representing the q-th turn of the k-th federal learning task for the task identification;

indicating the download transaction bill number, t_0→iIs a time stamp of the transaction bill; id_iIs the identity id of the participant;

is the content signed by the private key when the participant downloads; l is_iRepresenting a parameter location index; v_dowA credit amount representing a download cost; b is_iRepresenting the credit balance of the participant; b is₀Representing the current point balance of the uploading participant in the round;

s3-4: symmetric key EK is used by participants in the federal learning task^kDecrypting ciphertext E^k(W_new) Obtaining global model parameters W_new。

Images